Object Tracking With Dense Sampling Based On Spatio-Temporal Context And Kernelized Correlation Filters

Object tracking as an important research direction of computer vision has been widely used in security surveillance, motion analysis, activity recognition, intelligent transportation, human-computer interaction et al. In a recent study of the target tracking algorithm, tracking method based on detection treating the target tracking as an inspection tasks with evolution over time, online learning an discriminatory classifier which optimal separate of target from background, reflecting the changes of appearance model robustly, effectively coping with various challenging factors such as illumination, viewpoints, and shelter in the process of target tracking, has become the main trend of tracking method. In such method, tracking with dense sample which make full use of the sample in object and its surrounding caused wide public concern and one of the advantages of this method is it can use FFT to realize in real-time object tracking. The research object of this paper is base on two main tracking algorithms with dense sample spatio-temporal context and kernelized correlation filters and improved tracking method is propose aiming at the existing problems. The main contribution of this paper is as follows:(1) An improved multi-cue fusion tracking algorithm based on spatio-temporal context learning is proposed. The algorithm first got a new fast filter method through fusing the context learning and particle filter, which not only speeding up the filtering process but also improved the accuracy of the position predicted of object, even can track a large range of motion of object in two successive frames, at the same time, solved effectively the problem of estimating object scale. Secondly, to improve the tracking accuracy of the algorithm during tracking we used the observation model evaluating the possibilities of a variety of clues as object. Experiments resulting show that the average success rate of proposed improved algorithm increased by 13.51% on the basis of original spatio-temporal learning algorithm without considering the target scale changes and can better estimate the scale change of object.(2) An improved kernelized correlation filters tracking algorithm with prior constraint is proposed. Firstly, we established a discriminate classifier learning model based on priori constraint when use object model to learn the classifier of current frame adding tracking result of prior frame as constraint, using the model can enhance robust of the learned discriminate classifier for drastic changes of object appearance. Secondly, we considering the original kernelized correlation filters tracking algorithm only using gradient histogram for object appearance, did not make full use of the color of the target information. In improved algorithm, fused the mean color information into object representation which enhancing the capacity of the description of the object. Experiments resulting show that the average success rate of proposed improved algorithm increased by 7.82% on the basis of original tracking with kernelized correlation filters.We improved two tracking algorithm based on dense sample, each with different characteristics and their range of applications. Experiments show that the improved multi-cue fusion tracking algorithm based on spatio-temporal context learning suiting to track with task of target itself existing various changes and the improved kernelized correlation filters tracking algorithm with prior constraint has generalization ability which can manage a variety of external challenges during tracking.